Model or dataset
tenstorrent/tt-metal avatar
tenstorrent/tt-metal

TT-Metal: benchmark tables pinned to release candidates, a T/S formula that misses one row, and a package called ttnn

TT-NN operator library, and TT-Metalium low level kernel programming model.

127 stars47 forksC++Apache-2.0

At a glance

What is it?
Tenstorrent's tt-metal is two projects in one tree, an operator library called TT-NN and a kernel programming model called TT-Metalium. The performance tables are its most useful page and its most dated one: every published number was measured on a release candidate or a dated dev tag, two of them cite commits in a vLLM fork the same page calls deprecated, and the measurement formula reproduces four of the five rows.
Who is it for?
Read the tt-metal performance tables as a record of particular builds rather than as current capability, because that is what they are: five rows, five different tagged commits, and the newest of them is from a release series fifteen minor versions behind the branch. That does not make them useless, because a reproducible number pinned to a tag is worth more than a rounded one with no provenance, but it changes what you can conclude.
Can I use it commercially?
Yes. Apache-2.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
Is it still maintained?
Yes. The repository received new commits within the last day.
What is it written in?
Mainly C++, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • tenstorrent/tt-metal127 ★ · -1,481
Star history
Repository2026-07-302026-11-15
tenstorrent/tt-metal1.6k127

-1,481 Stars in the last 108 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

Every published number was taken on a candidate or a dev tag

The featured model tables carry a release column for each row, and not one of them names a released version. Llama 3.3 70B at tensor parallel 32 on Galaxy was measured on v0.65.0-rc7. Qwen 2.5 7B at TP=2 on the n300 was measured on v0.62.0-rc35. Qwen 2.5 72B at TP=8 on QuietBox was measured on v0.62.0-rc25. Mixtral 8x7B at TP=8 on QuietBox was measured on v0.62.0-dev20251015, and both Whisper rows were measured on v0.65.0-dev20251208. The branch, meanwhile, is shipping dev tags for the 0.80 series, dated per day. So the most recent build with a published figure is roughly fifteen minor series behind the code, and the two oldest are release candidates for the 0.62 series. That gap is the single most important fact about the tables: they are a snapshot of five moments in the project's history, and each row is reproducible only by checking out the tag named beside it.

The stated T/S formula holds on four rows and misses the fifth

The metrics block defines things precisely: TTFT is the time in milliseconds to generate the first output token after input arrives, tokens per second per user is throughput after prefill and is calculated as one divided by inter-token latency, and total tokens per second is T/S = T/S/U multiplied by batch size. Multiply and check. Qwen 2.5 7B gives 22.1 times 32, which is 707.2, and 707.2 is what the table prints. Qwen 2.5 72B gives 15.4 times 32, which is 492.8, printed. Mixtral gives 24.9 times 32, which is 796.8, printed. The Llama row does not: 2268.8 divided by 32 is 70.9, not the 72.5 in the row, and 72.5 times 32 would be 2320. The gap is too wide to be rounding of a single decimal place, and the page offers no explanation, so either the displayed T/S/U or the total is carrying a number the other column does not agree with. Whisper, at batch 1, is self-consistent at 105.0 on both counts.

Most rows miss their target and one has no target at all

Each row pairs a measured T/S/U with a target column, and the comparison is unflattering for four of five. Llama 3.3 70B reaches 72.5 against a target of 80, which is about 91 percent. Qwen 2.5 7B reaches 22.1 against 30, about 74 percent. Qwen 2.5 72B reaches 15.4 against 20, about 77 percent. Mixtral 8x7B reaches 24.9 against 33, about 75 percent. The Whisper rows invert the pattern in both directions. On the n150 the figure is 105.0 against a target of 45, so more than double. On the p150, the Blackhole part, the figure is 263.4 and the target cell is empty. That p150 row is also the only Blackhole measurement in the tables, and its hardware link points at the Blackhole page while every other row points at Wormhole, QuietBox or Galaxy. Hardware that has been announced as under active software optimisation is the one row with no published goal to miss, and it is the fastest row on the page.

The vLLM columns cite commits in a fork the page calls deprecated

Three of the tables carry a second provenance column, a vLLM Tenstorrent repo release given as a bare commit hash pointing into the tenstorrent/vllm repository. Llama 3.3 70B cites 59be953, Qwen 2.5 7B cites ced0161 and Qwen 2.5 72B cites e7c329b. A paragraph above the tables says the previously maintained vLLM TT fork is deprecated and sends readers to a separate vLLM TT plugin repository for installation and environment creation. So the recommended integration path is a different repository, while the benchmark rows keep pointing at pinned commits inside the one being retired, and the two tables that have such a column are exactly the three transformer rows. Anyone reproducing a row needs both repositories checked out at specific commits, one of which belongs to a project the README describes as no longer maintained.

Daily dev tags are excluded from the version a wheel reports

The releases in this repository are dated dev tags, one per day, such as v0.80.0-dev20261002 and v0.80.0-dev20261001, and the version they carry comes from setuptools-scm. Its configuration is explicit about what it will not count. The tag_regex accepts a plain v1.2.3 or a v1.2.3-rc1 and nothing else, and the git describe command excludes anything matching *-dev*. The fallback when no usable tag is found is 0.0.0.dev0. The two rules together have a consequence worth knowing before you pin anything: a dev tag is not a version setuptools-scm can parse, because a four part dev suffix is outside the regex, so a build cut from the daily tags reports a distance from the last parseable tag rather than the dated identity the tag itself carries. It is also why the benchmark tables can cite v0.65.0-dev20251208 and remain human readable while a wheel built from the same tree reports something else entirely.

The Python package is named ttnn and installs exactly one script

The repository is tt-metal, the low-level programming model inside it is TT-Metalium, the operator library is TT-NN, and the Python distribution is none of those: pyproject.toml declares the project name as ttnn, with the description General compute framework for Tenstorrent devices. One console script is registered, `tt-run`, resolving to ttnn.distributed.ttrun:main, so a pip install gives you a launcher for distributed runs and nothing else. The dependency list is the part that will bite first, because numpy is held at >=1.24.4,<2 while pandas is asked for 2.0.3 or newer, so the pinned numpy ceiling and a modern pandas have to resolve together in whatever environment you build. The build backend is pinned exactly, setuptools 80.10.2 and setuptools-scm 8.1.0, and requires-python is 3.10. Cibuildwheel is configured to skip musllinux builds, so the published wheels are glibc targets.

lib64 on any 64-bit Linux that is not Debian, Arch or Alpine

setup.py reimplements a small piece of GNUInstallDirs rather than using it, in a function whose docstring says it is inspired by that logic. The default directory is lib. It becomes lib64 only on Linux, only when the machine reports 64bit architecture, and only when none of three marker files exist: /etc/debian_version, /etc/arch-release or /etc/alpine-release. That check is a filesystem probe rather than a distribution lookup, so any 64-bit Linux that is none of those three gets a wheel laid out under lib64, and any machine carrying one of those files stays on lib even if its packaging conventions differ. A module level flag, BUNDLE_SFPI, is set to False in the visible part of the file, and a helper expands brace patterns in globs before the file's own build logic continues past that point. The root also holds three shell entry points, build_metal.sh, install_dependencies.sh and create_venv.sh, which is the practical alternative to the install guide for anyone not using pip.

Black formats a path list with a duplicate, ruff formats everything

Two formatters and an import sorter are configured, with matching line lengths of 120. Ruff is set to 120 globally with no include filter, while black's include is a regular expression covering a specific set of directories, and tests/scripts appears twice in it. isort follows the black profile, skips __init__.py, and sets skip_gitignore, so it reads paths that git ignores. Two of those three tools are named inside their own config, which means a file outside black's include list is formatted by ruff and sorted by isort but never touched by black. The static analysis side is broader still: .clang-format, .clang-format-ignore, .clang-tidy and a .codechecker.json paired with a .codechecker.skiplist, plus .yamllint, .pre-commit-config.yaml and a .cycodeignore. Test configuration is split across three places, pytest.ini, conftest.py and pyproject.toml, which is a common source of the question of which one wins.

Editorial conclusion

Read the tt-metal performance tables as a record of particular builds rather than as current capability, because that is what they are: five rows, five different tagged commits, and the newest of them is from a release series fifteen minor versions behind the branch. That does not make them useless, because a reproducible number pinned to a tag is worth more than a rounded one with no provenance, but it changes what you can conclude. Nobody expecting the hardware to meet its own targets will find that here, since four of the five rows sit between 73 and 91 percent of target and the fifth is an audio model with no target at all. Before you plan around this stack, check three things: which release you are actually on, since the tables name versions rather than main; whether your workload fits the fixed 128 token input condition; and whether you need the vLLM path at all, since the fork those columns cite is deprecated in favour of a separate plugin repository.

Frequently asked questions

What is the Python package name for tenstorrent/tt-metal?

ttnn, not tt-metal, according to pyproject.toml. It requires Python 3.10 or newer, depends on numpy held below version 2, and registers a single console script named tt-run.

Which Tenstorrent hardware do the tt-metal benchmark numbers come from?

From the featured tables: Galaxy and n300 and QuietBox on Wormhole, plus n150 on Wormhole and p150 on Blackhole, at tensor parallel factors of 32, 2 and 8, all reported for an input sequence length of 128 tokens.

Do the tt-metal featured model benchmarks reach their targets?

Most rows fall short. Llama 3.3 70B reports 72.5 against a target of 80, Qwen 2.5 7B 22.1 against 30, Qwen 2.5 72B 15.4 against 20, Mixtral 8x7B 24.9 against 33. The Whisper row on n150 reports 105.0 against 45, and the p150 row has no target.

Is the Tenstorrent vLLM fork still maintained?

The README states the previously maintained vLLM TT fork is deprecated and points to the vLLM TT plugin for installation, while two benchmark tables still cite commit hashes inside the deprecated fork.

How are tenstorrent/tt-metal release tags named?

A dated dev tag per day, such as v0.80.0-dev20261002. The packaging configuration excludes dev tags when calculating the version, and its tag_regex accepts only a plain v1.2.3 or a v1.2.3-rc1 shape, with 0.0.0.dev0 as the fallback.

What are TT-NN and TT-Metalium in tenstorrent/tt-metal?

TT-NN is the Python and C++ neural network operator library, and TT-Metalium is the low-level programming model for kernel development on Tenstorrent hardware. Each links to its own programming guide and API reference.

Official sources

  1. Official documentation
  2. Official README
  3. Project repository
  4. Release notes
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/tenstorrent-tt-metal.svg)](https://hysenlabs.com/projects/tenstorrent-tt-metal)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/tenstorrent-tt-metal.svg)](https://hysenlabs.com/projects/tenstorrent-tt-metal)

Was this useful?